Fair Use After Warhol: A Practical Guide to the Four Factors in the Age of AI

Summary. Fair use is the most important and least predictable doctrine in American copyright law, and the Supreme Court's 2023 decision in Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith changed how the first factor works without changing a word of the statute. This guide explains the four statutory factors of 17 U.S.C. § 107 one at a time, in plain English, and shows how Warhol narrowed "transformative use" by refocusing courts on the specific use challenged rather than on the new work as a whole. It traces the doctrine from Justice Story's 1841 opinion in Folsom v. Marsh through Campbell, Harper & Row, Sony, Google v. Oracle, and the appropriation-art cases, then applies the framework to software, search, news, education, and parody. A long section addresses the question every technology company is now asking: whether training a generative AI model on copyrighted works is fair use, and what the 2025 district court decisions in Bartz v. Anthropic, Kadrey v. Meta, and Thomson Reuters v. Ross Intelligence actually held. The guide closes with a step-by-step method for running a fair use analysis before you copy, a worked example, litigation-posture advice on burdens and summary judgment, an FAQ, and related reading.

In 1981, a young photographer named Lynn Goldsmith shot Prince for Newsweek. Three years later, Vanity Fair licensed one of her frames for $400 as an "artist reference," and the artist turned out to be Andy Warhol. Warhol made a silkscreen from it. Then he made fifteen more. Goldsmith did not learn that the other fifteen existed until Prince died in 2016 and one of them, Orange Prince, appeared on the cover of a commemorative magazine. The Warhol Foundation had licensed it to Condé Nast for $10,000. Goldsmith got nothing.

The Supreme Court took the case and, in Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023), held that the first fair use factor favored Goldsmith. Not because Warhol's silkscreen was unoriginal. Not because Warhol was a copyist. Because the particular use at issue, licensing a portrait of Prince to a magazine to illustrate an article about Prince, was the same thing Goldsmith's photograph was for, and it was commercial.

That distinction is the single most consequential development in fair use law in a generation, and it is widely misunderstood. This guide explains what the four factors actually require, what Warhol did and did not change, and how to apply the framework to the problems people bring to a lawyer's office: a documentary that uses forty seconds of a song, a competitor's product manual quoted in a comparison chart, a search index built from a billion web pages, a machine learning model trained on the entire internet.

The short answer

Fair use is a statutory defense in 17 U.S.C. § 107 that permits unauthorized use of copyrighted material when four factors, weighed together in light of copyright's purpose, favor the user. Nothing is "automatically" fair use. There is no word count, no percentage, no "30 seconds is fine" rule, and no magic incantation about education or non-profit status.

After Warhol, the practical questions to ask in order are:

1. What exactly are you doing with the work, and why do you need this particular work to do it? A use that serves a genuinely different purpose from the original has a strong first factor. A use that substitutes for the original, or competes in the same market, does not, and calling it "transformative" will not save it.
2. Is your use commercial, and does the commercial character matter here? Commerciality is not fatal, but it counts against you, and it counts more when your purpose overlaps with the original's.
3. How much did you take, and did you need that much? Taking the "heart" of a work is expensive even when it is short.
4. Does your use, and uses like yours if they became widespread, harm the market for the original or its licensing markets? This is where most close cases are actually decided.

Everything else is elaboration. What follows is the elaboration, because the elaboration is where cases are won.

Where fair use came from, and why that history still matters

Fair use is older than the statute that codifies it. In Folsom v. Marsh, 9 F. Cas. 342 (C.C.D. Mass. 1841), Justice Joseph Story, sitting as a circuit judge, decided whether a two-volume biography of George Washington that reproduced 353 pages of Washington's letters from a twelve-volume collection infringed. He said no to the defendant, and in doing so wrote the sentence that has governed ever since: courts must "look to the nature and objects of the selections made, the quantity and value of the materials used, and the degree in which the use may prejudice the sale, or diminish the profits, or supersede the objects, of the original work."

Read that again and you will notice that Story listed three of the four modern factors in a single sentence, in roughly the modern order, 135 years before Congress wrote them down.

Congress codified the doctrine in the Copyright Act of 1976 at 17 U.S.C. § 107. The statute's opening words matter more than most people realize:

Notwithstanding the provisions of sections 106 and 106A, the fair use of a copyrighted work, including such use by reproduction in copies or phonorecords or by any other means specified by that section, for purposes such as criticism, comment, news reporting, teaching (including multiple copies for classroom use), scholarship, or research, is not an infringement of copyright.

Two features of that preamble get litigated constantly. First, the list of favored purposes is illustrative, not exhaustive; the phrase is "such as," and the House Report accompanying the Act made clear that Congress intended to restate the judicial doctrine rather than freeze it. Second, appearing on the list is not a free pass. A "news reporting" use lost in Harper & Row Publishers, Inc. v. Nation Enterprises, 471 U.S. 539 (1985). A "teaching" use lost in the coursepack cases. The preamble tells you where to look, not what to find.

The statute then lists the four factors that a court "shall include" in its analysis:

1. the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes;
2. the nature of the copyrighted work;
3. the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
4. the effect of the use upon the potential market for or value of the copyrighted work.

A final sentence, added in 1992 after the Second Circuit had suggested otherwise, provides that the fact a work is unpublished "shall not itself bar a finding of fair use if such finding is made upon consideration of all the above factors."

The factors are not elements. You do not need to win all four, and you do not win by counting. The Supreme Court has said repeatedly that the factors are to be "weighed together, in light of the purposes of copyright," Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569, 578 (1994), and that fair use "calls for case-by-case analysis" rather than "bright-line rules." That is honest, and it is also why clients hate the answer.

Factor one: the purpose and character of the use

This is the factor that decides most cases, and it is the factor Warhol rebuilt.

Transformative use before Warhol

The modern first-factor inquiry traces to Judge Pierre Leval's 1990 Harvard Law Review article, Toward a Fair Use Standard, 103 Harv. L. Rev. 1105 (1990), which proposed that the central question is whether the new work is "transformative," meaning it "adds something new, with a further purpose or different character, altering the first with new expression, meaning, or message." The Supreme Court adopted that formulation four years later in Campbell, where 2 Live Crew's raunchy rewrite of Roy Orbison's "Oh, Pretty Woman" was held potentially fair as parody.

Campbell made three moves that still control:

• Parody gets latitude that satire does not. A parody comments on the original and therefore needs to conjure it up. Satire uses a work to comment on something else and, because it does not need this particular work, must justify the borrowing. The line is fuzzy in practice, but the reason for it is sound: if you are mocking the song, you must quote the song; if you are mocking society, you could have written your own song.
• Commerciality is not presumptively unfair. The Court rejected the reading of Sony Corp. of America v. Universal City Studios, Inc., 464 U.S. 417 (1984), that treated every commercial use as presumptively unfair, calling that a misreading of Sony's language.
• Transformativeness and commerciality trade off. "The more transformative the new work, the less will be the significance of other factors, like commercialism, that may weigh against a finding of fair use." Campbell, 510 U.S. at 579.

For twenty-five years, lower courts, especially the Second Circuit, read that sentence expansively. Blanch v. Koons, 467 F.3d 244 (2d Cir. 2006), found Jeff Koons's use of a fashion photograph of legs transformative. Cariou v. Prince, 714 F.3d 694 (2d Cir. 2013), went further, holding that twenty-five of Richard Prince's appropriations of Patrick Cariou's Rastafarian photographs were transformative as a matter of law because they "manifest an entirely different aesthetic," and suggesting that the new work need not comment on the original at all. Cariou was the high-water mark. Practitioners started advising clients that "different aesthetic" was enough.

What Warhol actually held

Warhol pulled that back. The Court's opinion, by Justice Sotomayor, made four points that every practitioner should be able to recite:

One: the unit of analysis is the use, not the work. The Court repeatedly stressed that § 107 asks about "the purpose and character of the use," and that the challenged use in the case was AWF's 2016 licensing of Orange Prince to Condé Nast, not Warhol's 1984 act of creation and not the Prince Series as an artistic achievement. Goldsmith had not challenged the museum display or the sale of the original works. As the Court put it, "the same copying may be fair when used for one purpose but not another." 598 U.S. at 533.

Two: the first factor asks whether the use has a "further purpose or different character," and that is a matter of degree. The Court did not abolish transformative use. It held that "the degree of difference must be balanced against the commercial nature of the use." A use that is only modestly different in purpose and highly commercial will not clear the first factor.

Three: substitution is the core concern. Both Goldsmith's photograph and AWF's licensed silkscreen were used, and were licensed, to illustrate magazine stories about Prince. That shared purpose meant the works were substitutes in that market. "Lynn Goldsmith's original works, like those of other photographers, are entitled to copyright protection, even against famous artists."

Four: the new work's added meaning matters, but it is not self-justifying. The Court worried, sensibly, that if "adds new meaning" were enough, the derivative work right in 17 U.S.C. § 106(2) would collapse, because every derivative work adds something. Book-to-film adaptations, translations, and sequels all add meaning; they are also squarely within the copyright owner's exclusive rights.

Justice Gorsuch, joined by Justice Jackson, concurred to emphasize that the statute directs attention to the use, not the artist's subjective intent. Justice Kagan, joined by Chief Justice Roberts, dissented at length, arguing that the majority undervalued artistic transformation and would chill creativity. The dissent is worth reading for its rhetorical force, but it is the dissent.

What Warhol did not hold

Three misreadings circulate, and all three are wrong.

Warhol did not hold that Warhol's Prince Series infringed. The Court decided one factor as to one use. It expressly did not decide the other factors or the ultimate fair use question, and it did not address whether making the works in 1984 was fair.

Warhol did not abolish transformative use. It cabined it. A genuinely different purpose still wins the first factor, as Google LLC v. Oracle America, Inc., 593 U.S. 1 (2021), decided two years earlier, illustrates and as the Warhol majority took pains to reaffirm.

Warhol did not create a rule that commercial uses lose. It created a sliding scale in which commerciality bites harder as purpose-overlap increases.

How to use factor one now

Frame your client's use at the right level of specificity, and be honest about it. The question is not "is our app transformative" but "what is the purpose of the specific act of copying, and does the original serve that same purpose for the same audience?" Three concrete illustrations:

• A biography quoting a memoir to analyze the author's account has a different purpose (criticism) than the memoir (telling the story). Strong factor one.
• A stock-photo aggregator that displays full-resolution images so users need not license them from the photographer has the same purpose. Weak factor one, regardless of how the images are cropped or captioned.
• A security researcher who decompiles firmware to find a vulnerability has a different purpose from the firmware's (running a device). Strong factor one, and this is the reasoning of Sega Enterprises Ltd. v. Accolade, Inc., 977 F.2d 1510 (9th Cir. 1992), and Sony Computer Entertainment, Inc. v. Connectix Corp., 203 F.3d 596 (9th Cir. 2000), which remain good law after Warhol precisely because the copying served an entirely different end.

Factor two: the nature of the copyrighted work

This is the factor that rarely decides anything and is therefore the factor lawyers under-argue. It asks where the copied work sits relative to the core of copyright's protection. Two sub-questions:

Creative or factual? Works closer to the "core of intended copyright protection," such as novels, songs, and photographs, get more protection than compilations of facts, technical manuals, and news reports. Campbell, 510 U.S. at 586. This does not mean factual works are unprotected; it means the scope of protectable expression is thinner because facts themselves are never protected. Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991).

Published or unpublished? Copying from an unpublished work weighs against fair use because the author's right of first publication is significant. Harper & Row, 471 U.S. at 564. Congress later added the closing sentence of § 107 to keep this from being dispositive, but it still counts.

Factor two did real work in Google v. Oracle. The Court held that Sun's declaring code, the API "labels and organization" that let programmers call methods, sits "further than most computer programs" from the core of copyright because its value comes largely from the investment of third-party programmers who learned it, not from the creativity of the code itself. That framing helped Google, and it is a template worth borrowing whenever the copied material is functional, interoperability-driven, or conventional.

Factor three: the amount and substantiality of the portion used

The rule is easy to state and hard to apply: you may take what you need for a legitimate purpose, and no more.

Quantity matters, but only relative to the whole. In Google v. Oracle, Google copied approximately 11,500 lines of declaring code, which sounds enormous until you learn it was about 0.4 percent of the Java SE API's roughly 2.86 million lines. The Court treated the copying as "tethered to a valid, and copyright-permissible, purpose": reimplementing a familiar interface so programmers could use what they already knew.

Quality matters more. Harper & Row is the canonical example. The Nation took roughly 300 words from President Ford's 200,000-word memoir, less than two-tenths of one percent. It lost, because the 300 words were "essentially the heart of the book," the passages about the Nixon pardon that made the memoir newsworthy.

Taking the whole thing is not automatically fatal. Time-shifting an entire television program was fair in Sony. Reproducing entire images at thumbnail size for a search index was fair in Perfect 10, Inc. v. Amazon.com, Inc., 508 F.3d 1146 (9th Cir. 2007), and in Kelly v. Arriba Soft Corp., 336 F.3d 811 (9th Cir. 2003). Intermediate copying of an entire program to reverse engineer it was fair in Sega. In each case, the full-work copying was necessary to the transformative purpose and the output did not substitute for the original.

The practical drafting lesson: when you must take a lot, build a record explaining why that amount was necessary. Courts respond to engineering and editorial necessity. They do not respond to "we took everything because it was easier."

Factor four: the effect on the market

Factor four asks whether the use, and unrestricted uses like it, would cause substantial harm to the actual or potential market for the copyrighted work, including licensing markets. Campbell, 510 U.S. at 590.

Substitution is the paradigm. If a consumer would buy the original but for the accused use, factor four is over. This is why the coursepack cases came out the way they did, and why Fox News Network, LLC v. TVEyes, Inc., 883 F.3d 169 (2d Cir. 2018), held that a service letting subscribers watch and download ten-minute clips of Fox broadcasts was not fair: it "usurped" a market Fox could exploit.

Licensing markets count, but the reasoning is circular unless you are careful. Every unauthorized use deprives the owner of a hypothetical license fee, so if "lost licensing revenue" always counted, no use would ever be fair. Courts limit the inquiry to "traditional, reasonable, or likely to be developed markets," American Geophysical Union v. Texaco Inc., 60 F.3d 913, 930 (2d Cir. 1994), and Campbell held that the market for criticism does not count: "the unlikelihood that creators of imaginative works will license critical reviews or lampoons of their own productions removes such uses from the very notion of a potential licensing market." 510 U.S. at 592.

The market for derivative works counts. This was decisive in Warhol, where the record showed Goldsmith licensed her Prince photographs to magazines and AWF licensed a derivative to a magazine.

Aggregate effects count. The question is not just what this defendant did but what would happen if the practice "should become widespread." Harper & Row, 471 U.S. at 568.

Two recent cases show factor four doing the heavy lifting. In Hachette Book Group, Inc. v. Internet Archive, 115 F.4th 163 (2d Cir. 2024), the court held that the Internet Archive's "controlled digital lending" program, which scanned print books and lent the scans on a one-to-one owned-to-loaned ratio, was not fair use. The scans were straightforward substitutes for the publishers' e-book licenses to libraries, a market that plainly exists and is plainly harmed. And in Warhol itself, the Court's first-factor analysis was really a market analysis wearing a purpose costume: the two works competed for the same magazine-cover licensing dollar.

The factors in the wild: eight contexts and how they come out

Doctrine is easier to hold onto when you can see it applied. Here are the recurring fact patterns, with the reasoning that decides them.

Parody and satire

Parody targets the original; satire targets something else using the original as a vehicle. Campbell protects the first robustly. The second must justify itself. Dr. Seuss Enterprises, L.P. v. ComicMix LLC, 983 F.3d 443 (9th Cir. 2020), is the cautionary tale: a mashup of Oh, the Places You'll Go! with Star Trek was held not to be parody, because it did not comment on Seuss, and not fair use, because it copied heavily and targeted the same market. After Jack Daniel's Properties, Inc. v. VIP Products LLC, 599 U.S. 140 (2023), the parallel trademark analysis has also tightened; parody that functions as a source identifier gets no special First Amendment filter.

Documentary film and archival footage

Documentaries live on factor one. Footage used to illustrate the subject of commentary (showing the speech being analyzed) is far stronger than footage used to entertain (a montage set to music). Duration matters less than function. The Documentary Filmmakers' Statement of Best Practices in Fair Use, developed at American University, is not law, but courts and insurers treat industry norms as evidence of reasonableness, and errors-and-omissions carriers will ask about it.

Photographs of copyrighted works

Photographing a sculpture or a building raises its own rules. Architectural works publicly visible from a public place are exempted by 17 U.S.C. § 120(a), so photographing a building is not infringement at all, no fair use analysis required. Sculptures are different; there, fair use does the work, and incidental or background inclusion generally wins on factors three and four.

Software: reverse engineering and interoperability

Sega, Connectix, and Google v. Oracle form a coherent line. Intermediate copying to study a program, extract unprotected functional elements, and build an interoperable product is generally fair, provided (a) the copying was necessary because the functional information was not otherwise available, (b) the copier had a legitimate reason for access, and (c) the final product does not contain the copied expression. Note the sequencing: the fair use defense protects the intermediate copy, not the shipped one.

The engineering point worth emphasizing to a technical audience: what makes API declaring code special is that its expressive content is minimal and its network value is enormous. The java.lang.Math.max(int, int) signature is not where Sun's creativity lived. Where Sun invested creativity, the implementing code, Google wrote from scratch. Courts increasingly understand that distinction, and briefs that explain it clearly do better than briefs that gesture at "interoperability."

Search engines and indexes

Building an index requires copying everything. Courts have consistently held that the copying is fair when the output is a pointer rather than a substitute. Kelly v. Arriba Soft, Perfect 10 v. Amazon, and Authors Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015), all turn on the same insight: a thumbnail or a snippet sends users to the original rather than replacing it. Authors Guild is the most useful of the three for modern disputes because Judge Leval, who invented the transformative-use framing, wrote it, and because he was careful to say that the snippet function was designed with "blacklisting" and length caps precisely to prevent substitution. Design choices that limit substitution are evidence.

Education and classroom use

The preamble lists teaching, and § 110(1) separately exempts in-classroom performances and displays. But copying course materials in bulk is a different matter. Princeton University Press v. Michigan Document Services, Inc., 99 F.3d 1381 (6th Cir. 1996) (en banc), held that a copy shop's coursepacks were not fair use, largely on factor four, because a functioning permissions market existed. The lesson for universities is that "educational" is a thumb on the scale, not a shield, and that the existence of an easy licensing mechanism is what kills these cases.

News reporting and the scoop problem

Harper & Row remains the governing decision: scooping an unpublished work's most newsworthy content is not fair, even by a news organization, even in small quantity. But routine quotation for reporting and commentary is core fair use. The distinction is between reporting about a work and appropriating the value of the work.

Appropriation art after Warhol

The honest post-Warhol advice to artists and galleries is that "different aesthetic" alone is no longer enough when the new work is licensed into the same market as the source. Artists whose work comments on the source, or who use the source in a medium and market entirely unlike the original's, remain in reasonable shape. Artists who license appropriated images for commercial products in the same channels as the source photographer face real exposure. Insurance underwriters have noticed.

Fair use and generative AI

No fair use question has ever arrived this fast or this large. Nearly every foundation model is trained on copyrighted works, and dozens of suits have been filed since late 2022. As of this writing, several district courts have ruled, no federal court of appeals has decided the training question, and the Supreme Court has not taken a case. Anyone who tells you the answer is settled is selling something.

Separate the three questions

Most of the confusion in public debate comes from mashing three distinct legal questions together. Keep them apart:

1. Acquisition. How did the training corpus get assembled? Downloading a pirate library is a different act from buying books, scanning them, and discarding the paper. This question is about the copy made to obtain the work, and it is often the defendant's biggest problem.
2. Training. Does the act of processing works to adjust model weights infringe, and if so is it fair use? This is the transformative-purpose question.
3. Output. Does a particular generation reproduce protected expression, and if so who is liable? This is ordinary substantial-similarity analysis plus, sometimes, contributory liability.

What the 2025 decisions actually held

Training on lawfully obtained books was held fair use, but retaining pirated copies was not. In Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA (N.D. Cal. June 23, 2025), Judge Alsup granted partial summary judgment holding that using lawfully purchased books to train a large language model was "exceedingly transformative" and fair, and that scanning purchased print books to create digital copies for internal use was a permissible format shift. He separately held that downloading and retaining a library of pirated books was not fair use, because the acquisition served no transformative purpose independent of training and because a market for licensed copies exists. The case later settled, with a reported $1.5 billion class settlement announced in September 2025, which is itself a data point about how defendants price acquisition risk.

A second court reached a similar training result on a different record and flagged a new theory. In Kadrey v. Meta Platforms, Inc., No. 3:23-cv-03417-VC (N.D. Cal. June 25, 2025), Judge Chhabria granted partial summary judgment to Meta on fair use as to the thirteen plaintiff authors, while writing at length that the plaintiffs had simply failed to develop the evidentiary record on the theory most likely to succeed: market dilution, the idea that a model capable of generating an effectively unlimited supply of competing works in a genre harms the market for the training works even if it never reproduces any of them. He was explicit that his ruling did not mean training is generally lawful. Practitioners should read that opinion as a roadmap being handed to the next plaintiff.

Training on a competitor's editorial content to build a competing product was not fair use. In Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-cv-00613-SB (D. Del. Feb. 11, 2025), Judge Bibas, sitting by designation, revised his earlier opinion and granted summary judgment to Thomson Reuters on fair use as to Westlaw headnotes used to train a legal research tool. Factors one and four drove it: Ross was building a substitute for Westlaw, and the use was commercial and non-transformative in purpose. The case is the cleanest illustration of the Warhol principle applied to AI. The technology was novel; the purpose was competitive substitution.

How the four factors line up for model training

Factor one. Defendants argue that training extracts statistical relationships, not expression, and serves an entirely different purpose from the works themselves. That is a serious argument, and Bartz and Kadrey accepted it. Plaintiffs answer with Warhol: look at the specific use and ask whether the model competes in the same market as the works. Where the model is designed to produce the same kind of output as the training works (a novel-generating model trained on novels, a legal-research tool trained on legal research), the purpose overlap is real.

Factor two. Training corpora are heavy on expressive works, which favors plaintiffs, though the defense response is that the model learns unprotectable facts, style, and structure, none of which copyright covers. Feist and the idea/expression dichotomy of § 102(b) do real work here.

Factor three. Models are trained on complete works. Defendants rely on the search-index cases, where whole-work copying was necessary and the output was non-substitutive. Plaintiffs rely on cases where whole-work copying was fatal because it was not necessary.

Factor four. This is the battlefield. Three theories are in play. Direct substitution (the model reproduces the work, or produces outputs users buy instead of the work) is the strongest but hardest to prove and is often limited to memorized passages. Lost licensing market (a market for AI training licenses now demonstrably exists, given the deals major publishers have signed) is powerful but faces the circularity objection. Market dilution, the theory Judge Chhabria described, is the most consequential if it takes hold, because it does not require any reproduction at all.

What to advise a client building or using a model

• Provenance is the whole ballgame on acquisition. Keep records showing how every corpus was obtained. "We do not know where the scrape came from" is the worst possible answer.
• Contract risk is separate from copyright risk. A website's terms of service, a database license, or a robots directive can create breach-of-contract and, in some jurisdictions, related claims that fair use does not touch. See our discussion of scraping in Data Scraping After hiQ v. LinkedIn.
• Engineer against memorization and against substitution. Deduplication, filtering of near-verbatim outputs, and retrieval designs that cite and link to sources are not just good practice; they are factor-four evidence.
• Indemnities are the market's actual answer. Enterprise buyers now demand IP indemnities for model outputs. Read the caps, the carve-outs for customer prompts, and the duty-to-defend language carefully. See Indemnification and Limitation of Liability.
• Registration still matters on the plaintiff side. Statutory damages and fees require timely registration under 17 U.S.C. § 412, which is why so many AI complaints are framed around registered works. See Copyright Damages and Remedies.

For the broader picture of how these suits are structured and what relief plaintiffs are seeking, see Copyright Infringement Claims Against Generative AI.

A method: how to actually run a fair use analysis

Here is the sequence I use, and the sequence I recommend to in-house teams building a repeatable process.

Step 1. Confirm there is a copyright and that it covers what you want. Facts, ideas, procedures, and systems are outside copyright under 17 U.S.C. § 102(b). Short phrases, titles, and blank forms are not protected. Federal government works are in the public domain under § 105. If the material is unprotected, you never reach fair use, and saying "fair use" out loud when you did not need it is a gift to the other side.

Step 2. Check whether a specific exemption applies. Fair use is the general safety valve, but the Act contains many specific ones: § 108 (libraries and archives), § 109 (first sale), § 110 (classroom and religious performances), § 117 (computer program copies and adaptations by owners), § 120 (architectural works), § 121 (accessible formats). A specific exemption is a cleaner defense than a four-factor argument.

Step 3. Define the use with precision. Write one sentence: "We will copy X, in quantity Y, in order to Z, and distribute it to audience A through channel B." If you cannot write that sentence, you are not ready to analyze.

Step 4. Run each factor and write down the answer. Not in your head. Contemporaneous written analysis is evidence of good faith, and while good faith is not a factor, courts notice. Note that this document may be privileged, and should be routed through counsel if you want it to stay that way. See Attorney-Client Privilege and Work Product for Businesses.

Step 5. Stress-test factor four with the widespread-use question. If everyone did this, what happens to the market for the original? If the answer is "it collapses," reconsider.

Step 6. Reduce the take. Almost every fair use analysis improves if you use less, use lower resolution, link instead of copy, or add commentary that makes the purpose visible on the face of the work.

Step 7. Decide whether to license anyway. Fair use is a defense, which means you assert it after you have been sued. A $2,000 license is cheaper than a $200,000 defense you win.

A worked example

Northbeam Analytics, Inc. is a fictional startup that sells a competitive-intelligence dashboard to retail chains. It wants to build a feature that ingests competitors' weekly circulars, which are PDF documents containing photographs, prices, and marketing copy, and shows clients a normalized price comparison. Three design options are on the table.

Option A: display the circular pages in the dashboard. Users see the competitor's PDF rendered in full. Factor one is weak: the purpose (letting retail buyers look at circulars) is the same as the circular's purpose. Factor three is at maximum. Factor four is bad, because the publishers license circular data commercially. This option is the TVEyes problem, and the answer is no.

Option B: extract prices and product names into a table, and show no images or marketing copy. Now Northbeam is copying facts. Prices are facts. Product names are typically unprotected short phrases and may implicate trademark, not copyright. The extraction step involves an intermediate copy of the whole PDF, which looks like Sega: whole-work copying necessary to reach unprotected functional information, with no protected expression in the output. Factor one is strong, factor two favors Northbeam because the protectable expression is thin, factor three is justified by necessity, and factor four is strong because a price table does not substitute for a circular. This option is defensible.

Option C: show a thumbnail of the circular page next to each extracted row, linking to the competitor's site. Now we are in Perfect 10 and Authors Guild territory. The thumbnail is a pointer, not a substitute, and it serves an identification function the price table alone cannot. It is a closer call than Option B but a reasonable position, and the risk drops further if the thumbnail is low-resolution, small, and clickable through to the source.

Notice what did the work in this example: not the label "transformative," but a concrete account of what is copied, why, and what the output substitutes for. That is how the analysis should read in a memo.

What fair use is not

Fair use is not permission. It is an affirmative defense. You raise it after you are sued, you carry the burden of proving it, Campbell, 510 U.S. at 590, and you must plead it under Fed. R. Civ. P. 8(c).

Fair use is not a defense to breach of contract. If you agreed to terms that prohibit the use, fair use does not rescue you from the contract claim, though preemption arguments under 17 U.S.C. § 301 sometimes get traction where the contract claim is really a copyright claim in disguise.

Fair use does not defeat a § 1201 anti-circumvention claim. Breaking a technological protection measure violates 17 U.S.C. § 1201(a)(1) even if what you then do with the work is fair use. See our full treatment in The DMCA Anti-Circumvention Provisions.

Fair use is not an excuse to skip attribution, and attribution is not fair use. Crediting the author does not make an unfair use fair, and failing to credit does not make a fair use unfair, though removing copyright management information can create separate liability under 17 U.S.C. § 1202. Murphy v. Millennium Radio Group LLC, 650 F.3d 295 (3d Cir. 2011).

Fair use matters on the takedown side. A copyright owner sending a DMCA notice must consider fair use in good faith before sending it, or risk liability under 17 U.S.C. § 512(f). Lenz v. Universal Music Corp., 815 F.3d 1145 (9th Cir. 2016). See How to File a DMCA Takedown Notice and Respond to One.

Litigation posture: how fair use is actually decided

Who decides. Fair use is a mixed question of law and fact. Google v. Oracle held that where a jury has resolved the underlying historical facts, the ultimate fair use determination is a legal question for the judge, reviewed de novo on appeal, with deference to the jury's subsidiary findings. That holding matters enormously for trial strategy: it means a defense verdict on fair use can be reversed as a matter of law, and it means the jury instructions and special verdict form deserve more attention than they usually get.

Summary judgment. Fair use is frequently resolved on summary judgment when the facts about the use are undisputed, which they often are. Both Warhol and Google arrived at the Supreme Court from summary judgment postures. The practical implication is that the record built in discovery, particularly documents about why the defendant chose this work and what markets the plaintiff licenses into, decides the case. See Summary Judgment Under Rule 56.

Discovery targets. Plaintiffs should seek documents on the defendant's licensing consideration, internal discussions of alternatives, and revenue attributable to the use. Defendants should seek the plaintiff's actual licensing history, evidence about whether the claimed market exists in fact, and any evidence that the plaintiff has tolerated similar uses. Our Discovery Toolkit covers the mechanics.

Fees. A prevailing party on either side may recover attorney's fees under 17 U.S.C. § 505, and the standard is not one-sided: Fogerty v. Fantasy, Inc., 510 U.S. 517 (1994), and Kirtsaeng v. John Wiley & Sons, Inc., 579 U.S. 197 (2016), instruct courts to give substantial weight to the objective reasonableness of the losing party's position, while considering all the circumstances. A defendant with a colorable fair use position who wins is in a good position to be made whole. A plaintiff who sues over a plainly fair use may pay for the privilege.

Quick reference: how the leading cases came out

A pre-copy checklist

• Identify precisely what will be copied, in what quantity, and in what form.
• Confirm the material is actually protected (not facts, not ideas, not a short phrase, not a U.S. government work, not in the public domain).
• Check whether a specific statutory exemption applies before reaching for § 107.
• Write the one-sentence use statement (copy X, quantity Y, purpose Z, audience A, channel B).
• Ask whether the original serves the same purpose for the same audience. If yes, stop and reconsider.
• Ask whether you need this work or merely a work. Satire-style borrowing needs a better reason than parody-style borrowing.
• Reduce the amount to what the purpose requires, and document why that amount was necessary.
• Identify the plaintiff's actual and likely licensing markets, and ask whether widespread copying like yours would damage them.
• Consider design mitigations: lower resolution, snippets, links to the source, caps on retrieval, output filtering.
• Record provenance for every source, especially for datasets.
• Route the analysis through counsel so it can be privileged.
• Price a license and compare it honestly to the cost of a defense.

Frequently asked questions

Is there a percentage of a work I can safely use? No. There has never been one. Harper & Row lost over 0.15 percent of a book; Sony won on 100 percent of a television program. The percentage tells you almost nothing without the purpose and the market analysis.

Does giving credit make it fair use? No. Attribution is good practice and it matters to academic and journalistic ethics, but it is not one of the four factors. Removing an author's credit can create a separate claim under 17 U.S.C. § 1202, so attribution can help you avoid a different problem while doing nothing for your fair use case.

I am a nonprofit. Doesn't that settle it? It helps on factor one, and it is expressly mentioned in the statute. It does not settle anything. The coursepack and photocopying cases involved educational and research uses that still lost on factor four.

Does adding a filter, changing the colors, or "making it my own" create a transformative use? After Warhol, be careful. A different visual style is a difference in character, and it counts, but the Court balanced that difference against commerciality and asked whether the use served a different purpose. Style changes that leave the work serving the same function for the same audience are weak.

Can I use a song in my YouTube video if it is under 30 seconds? There is no 30-second rule; it is folklore. Whether a short musical use is fair depends on the same four factors, and music has robust licensing markets that make factor four difficult. Separately, platform content-identification systems operate on private license agreements, not on copyright law, so a use that would be fair in court can still be claimed or blocked on a platform. Your remedy there is the platform's dispute process, not litigation. See Digital Millennium Copyright Act Safe Harbors for Online Service Providers.

If I win on fair use, do I get my fees? Possibly. Section 505 makes fees discretionary, and Fogerty and Kirtsaeng direct courts to weigh the objective reasonableness of the loser's position heavily. Prevailing defendants recover fees more often in copyright cases than in most areas of American litigation.

Does fair use exist outside the United States? Not in this form. Most countries use "fair dealing" or enumerated exceptions, which list specific permitted purposes rather than applying an open-ended balancing test. A use that is fair in the United States may infringe in the United Kingdom, Germany, or Japan. Cross-border products need a jurisdiction-by-jurisdiction analysis; see Global Patent Litigation Strategies for how similar cross-border problems get managed in the patent context.

Is training an AI model on my company's licensed data safe? Safer, but read the license. Many content licenses were written before model training existed and grant rights for "internal business purposes" or "analysis" without contemplating derived model weights. The question of who owns and may use the model that results is a contract question first and a copyright question second. See Drafting Software License Agreements.

What should I do if I receive a demand letter over a use I think is fair? Do not delete anything; a preservation obligation attaches when litigation is reasonably anticipated. See Litigation Holds, Spoliation, and Rule 37(e). Then get the analysis done properly and respond. Many fair use disputes end with a modest license or a design change, and the leverage in that negotiation comes from a defense that is actually written down.

Closing thought

The best description of fair use I know is that it is copyright's built-in safety valve for uses that copyright's ordinary rules would foolishly forbid. Justice Story understood that in 1841, Congress wrote it down in 1976, and the Supreme Court has spent fifty years reminding everyone that it resists mechanical application.

Warhol did not break the doctrine. It corrected a drift. For twenty years, "transformative" had been doing more work than any single word can bear, and courts were finding transformation in changes of style that left the accused use competing head-to-head with the original. The Court's answer was to send everyone back to the statute's actual text, which asks about the use, and to the doctrine's actual purpose, which is to permit the uses that copyright's monopoly should never have reached.

For clients, the operational lesson is unglamorous but reliable: describe the use precisely, take only what the purpose requires, build the product so it points at the original rather than replacing it, keep records of where your material came from, and get the analysis in writing before you ship. Do those five things and most fair use problems either disappear or become manageable.

Related articles

• Copyright Overview — what copyright protects, how long it lasts, and the exclusive rights fair use limits.
• Copyright Damages and Remedies: Statutory Damages, Profits, Injunctions, and Fees — what is at stake when a fair use defense fails.
• The DMCA Anti-Circumvention Provisions — why fair use does not answer a § 1201 claim.
• Copyright Infringement Claims Against Generative AI — the pending AI cases and what plaintiffs are seeking.
• How to File a DMCA Takedown Notice and Respond to One — the § 512 process and the § 512(f) fair use duty.
• Copyright Registration: A Comprehensive Guide — why registration timing controls remedies.
• Data Scraping After hiQ v. LinkedIn — contract and CFAA claims that fair use does not reach.
• Legal Protection of Software: Copyrights, Patents, Trade Secrets, and Contracts — where the reverse-engineering cases fit.
• Summary Judgment Under Rule 56 — the posture in which most fair use cases are decided.
• Artificial Intelligence Key Legal Issues: A Comprehensive Overview — the wider AI risk picture for businesses.

This article is provided for general informational purposes and does not constitute legal advice. Fair use is intensely fact-specific and the law governing AI training is developing rapidly; the analysis above reflects authority available as of the date last modified. Consult qualified copyright counsel about any particular use.